Skip to main content Skip to secondary navigation
Page Content

A Data-Driven Approach to Understanding How the Brain Works

A meta-analysis of 18,000 fMRI studies challenges neuroscientists’ understanding of brain functions and reaffirms the need for more targeted treatments for mental disorders.

an image of a brain with different parts lighted up in vibrant colors

A recent Stanford study shows that different parts of the brain work together in surprising ways that differ from current neuroscientific wisdom. | iStock/dakuk

People are forever categorizing things based on superficial traits only to discover that, upon closer examination, those groupings don’t hold. Take an example from the produce aisle: Yams and sweet potatoes look and taste similar, but biologists know they come from completely unrelated plants. Meanwhile, kale, cauliflower, broccoli, and brussels sprouts seem very different from one another, but in fact they are the same species. 

These same sorts of mistaken impressions have arisen in neuroscience: We’ve lumped and split different types of mental phenomena based on long-held beliefs derived from psychology, but when we look at the data from brain scans, the categories we’ve imagined might exist aren’t always grounded in biological reality, says Ellie Beam, an MD/PhD candidate at Stanford. Issues of categorization have been particularly problematic for mental disorders: Our definitions are based on symptoms and don’t map well onto data from brain scans. 

And the ways we categorize how brain functions map onto brain structures matter: Research is proposed, funded, and organized around our existing theories about what’s commonly known in neuroscientific circles as “functional domains,” that is: which brain structures and collections of structures are responsible for which brain functions.

Beam wondered whether our current system for categorizing functional domains could be improved upon if the data could speak for itself. Would a data-driven categorization of brain functional domains better align with the neuroimaging data? 

To answer that question, Beam and her colleagues used natural language processing and machine learning to analyze more than 18,000 research papers that contained results from brain scans gathered primarily using functional magnetic resonance imaging (fMRI) technology. These scans study which parts of the brain leap into action when people are asked to do various things such as speak, move, rest, think logically, remember events from the past, or feel particular emotions. 

Read the full study, "A Data-driven Framework for Mapping Domains of Human Neurobiology"


The team extracted and clustered more than 1,600 brain function terms (like memory, reward, cognition, emotion, or arousal) found in the fMRI papers and then mapped those terms to the brain circuits (groupings of brain structures) highlighted in the same papers. This analysis showed that different parts of the brain work together in surprising ways that differ from current neuroscientific wisdom. In particular, the study calls into question our current understanding of how brains process emotion. Beam’s work could lead to a better understanding of mental disorders and, eventually, to more successful treatments. 

“When we use the data to tell us what a functional domain is, it can surprise us,” Beam says. “And that’s because the brain is doing something that we didn’t realize it was doing when we started this whole mapping endeavor.” The work was recently published in Nature Neuroscience.

images of brain scans with the different domains and feelings associated with those domains listed

The Stanford team’s approach tied mental function terms used in fMRI research papers to specific brain circuits (colored in the brain images), yielding six functional domains that differ from conventional wisdom.

What’s in a Domain?

For more than 30 years, researchers have been using fMRI to measure how blood ebbs and flows across the brain, clueing us in to which parts are more active when people are asked to do various activities. Over the years, the fMRI mapping of various brain functions to specific groupings of brain structures has yielded an entirely new way to categorize basic brain activity. Indeed, in 2009, the National Institute of Mental Health developed a system called the Research Domain Criteria (RDoC) project that proposes groupings of brain functions and major portions of the brain based on experts’ assessments of all the neuroscientific literature, including fMRIs. 

According to the RDoC system, there are six major “domains” of brain function, each of which is a set of systems. For example, our sensorimotor systems that allow us to taste, smell, see, and hear as we move around the world are categorized as a functional domain. And our cognitive systems, including attention, perception, language, and working memory, constitute a domain. Arousal is another domain, related to sleep and wakefulness and the whole continuum of sensitivity to stimuli. There is also a social processes domain that concerns how we navigate our relationships with others. And the experts have carved out two “valence” domains. One for “positive valence systems” that involve motivational thinking such as being goal-directed or reward-seeking, and another for “negative valence systems” that involve responding to attacks or threats of harm with emotions such as fear or anxiety. 

The Diagnostic and Statistical Manual (DSM), the reference used by psychiatrists, predates the RDoC and offers an alternative categorization of brain function derived from our understanding of mental disorders. Thus, the DSM defines many types of mental disorders including depressive, anxiety, trauma- and stressor-related, obsessive-compulsive, disruptive, substance abuse, neurodevelopmental, psychotic, and bipolar. 

In fMRI studies, researchers have been stymied in their efforts to identify brain circuits that are closely tied to the DSM categories. Indeed, some disorders seem to be defined too broadly. They are like sweet potatoes and yams. Their symptoms seem the same, so they receive the same diagnosis, but look a little deeper and they don’t exhibit a shared set of changes to their brain circuitry. Meanwhile, other disorders seem to be defined too narrowly. Like kale and cauliflower, they seem different on the surface, but dig more deeply and you’ll find similarities in their brain scans. 

“If the goal is to develop biologically based treatments for mental health problems, we need to start by better characterizing how circuits are functioning in individuals rather than focusing on what their symptoms are,” Beam says.

A Data-Driven Approach Yields Surprises

For her research, Beam and her colleagues set out to discover whether data from fMRI papers would deliver a new way of categorizing functional domains in healthy as well as disordered brains. They started with more than 18,000 fMRI research papers published over a 25-year period, which they divided into training and test sets for purposes of validation.

Beam first extracted the x, y, z coordinates of brain areas described in each paper and mapped them onto 118 gray matter structures in a standardized brain atlas. Often the coordinates associated with a particular function include multiple structures combined into what are called circuits. 

Beam and her colleagues next extracted more than 1,600 brain function terms – including things like memory, reward, cognition, valence, emotion, or arousal – from publicly available sources (including the RDoC) and determined where those terms co-occurred in the same paper with particular circuits. This allowed them to map function terms to specific circuits in the brain atlas. After a series of clustering steps, the distinctive sets of mental function terms associated with a circuit were narrowed down to the 25 or fewer most salient, which were used to name the data-driven domains of brain function. 

Beam settled on using six data-driven domains for purposes of comparison with the RDoC and DSM. (Readers can explore the alternatives at But in order to make that comparison, the team had to first map function words associated with these two systems onto the same brain circuitry they had extracted from the 18,000 fMRI papers. As expected, the DSM’s disorders didn’t map onto the brain atlas in a coherent way. And although the RDoC domains mapped fairly well onto the brain atlas, they were still less consistently well matched to particular circuits than were the data-driven domains.

The six data-driven domains – memory, reward, cognition, vision, manipulation, and language – differed from the RDoC and DSM categories in a variety of ways that merit further examination by neuroscientists. But even to a layperson it’s plain to see that none of these relate to emotion. By contrast, according to the RDoC, the brain has a distinct domain for processing emotional intensity (arousal) and two distinct domains dealing with valence (positivity or negativity). In the data-driven domains, however, these features are integrated into circuits for memory and reward, which is consistent with some recent theories that emotions are constructed and aren’t located in particular brain circuits. 

In addition, the mapping of the data-driven domain for cognition included certain emotional terms and two structures – the insula and the cingulate – that are thought to be areas of common vulnerability across mental disorders. “If certain parts of the brain like the insula and the cingulate are supporting both cognitive and emotional processes, that violates the idea that we have from psychology that cognitive and emotional domains are separate.” 

Lessons Learned for Psychiatry

Beam hopes that studies like hers can help researchers move away from the disease model of mental disorders and toward characterizing patients according to how their circuits are functioning. “This way of thinking could eventually lead to better biologically based treatments,” she says. “Perhaps patients’ mental health will even be expressed in terms of the circuits that we can change rather than as a single diagnostic label.”

Stanford HAI's mission is to advance AI research, education, policy and practice to improve the human condition. Learn more

More News Topics